Models · speech to text
Whisper
Whisper is OpenAI’s speech recognition model. You give it audio and it gives you text. It is not a chat model. It does not answer questions.
What it is
The multilingual sizes can also translate speech in other languages into English. OpenAI’s GitHub README says the code and the model weights are released under the MIT License. The license line on each size page is “MIT, per OpenAI.” Hugging Face lists apache-2.0 on most of these repos, and mit on large-v3-turbo. I follow the GitHub statement and keep that mismatch in the footnote.
Released: September 2022 (original), December 2022 (large-v2), November 2023 (large-v3), September 2024 (large-v3-turbo), per model-card.md. Source: model-card.md.Audio in 30-second windows (transcribe() slides across longer files). Paper: arXiv 2212.04356.
large-v1 (OpenAI's original 'large', Sept 2022) and the distil-whisper checkpoints are mentioned on the family page but not given their own entries.
The sizes
Parameter counts, “Required VRAM,” and whisper.cpp file sizes are Vendor-reported from OpenAI’s README and the whisper.cpp model list. “Required VRAM” is OpenAI’s figure for its own Python package. FP16 weights are Computed: parameter count times 2 bytes, from the official safetensors header on 19 Nov 2026. A speed cell is Measured only when the 19 Nov run included that file. Otherwise it is Not yet measured.
| Size | English-only file | Parameters | OpenAI VRAM | whisper.cpp file | whisper.cpp memory | FP16 weights | 161 s clip |
|---|---|---|---|---|---|---|---|
| tiny | tiny.en | 39 M Vendor-reported | ~1 GB | 75 MiB | ~273 MB | 0.07 GiB Computed | 21.68× Measured |
| base | base.en | 74 M Vendor-reported | ~1 GB | 142 MiB | ~388 MB | 0.14 GiB Computed | 13.68× Measured |
| small | small.en | 244 M Vendor-reported | ~2 GB | 466 MiB | ~852 MB | 0.45 GiB Computed | 5.74× Measured |
| medium | medium.en | 769 M Vendor-reported | ~5 GB | 1.5 GiB | ~2.1 GB | 1.42 GiB Computed | 2.66× Measured |
| large-v2 | none | 1550 M Vendor-reported | ~10 GB | 2.9 GiB | ~3.9 GB | 2.87 GiB Computed | Not yet measured |
| large-v3 | none | 1550 M Vendor-reported | ~10 GB | 2.9 GiB | ~3.9 GB | 2.87 GiB Computed | Not yet measured |
| large-v3-turbo | none | 809 M Vendor-reported | ~6 GB | 1.5 GiB | not published for turboNot yet measured | 1.51 GiB Computed | 2.17× Measured |
Quantized whisper.cpp files, Vendor-reported: large-v2-q5_0 and large-v3-q5_0 are 1.1 GiB. large-v3-turbo-q5_0 is 547 MiB. The q5_0 timing lives on the turbo page, not on its own page. OpenAI’s relative speeds (tiny ~10×, base ~7×, small ~4×, medium ~2×, turbo ~8×, large 1×) are from an A100 with English audio. That is Vendor-reported and it is not this machine.
Measured on Frankenstein (HP ProLiant DL380p Gen8), 19 Nov 2026
whisper.cpp commit ca19af24, CPU only, 2× Xeon E5-2697 v2, 24 threads, English. The longer clip is the LibriVox reading of the Gettysburg Address, public domain. The short clip is whisper.cpp’s own JFK sample. Both are in the table.
| Model | Audio | File on disk | Wall time | Audio ÷ wall | Encoder | Peak RSS | JFK word errors |
|---|---|---|---|---|---|---|---|
| tiny | JFK sample, 11.0 s | 74 MiB | 0.81 s | 13.50× | 0.25 s | 172 MiB | 0.0% (0/22 words) |
| tiny | LibriVox Gettysburg Address, 161.2 s | 74 MiB | 7.44 s | 21.68× | 1.41 s | 271 MiB | — |
| base | JFK sample, 11.0 s | 141 MiB | 1.32 s | 8.33× | 0.47 s | 279 MiB | 0.0% (0/22 words) |
| base | LibriVox Gettysburg Address, 161.2 s | 141 MiB | 11.79 s | 13.68× | 2.87 s | 379 MiB | — |
| small | JFK sample, 11.0 s | 465 MiB | 3.29 s | 3.35× | 1.49 s | 745 MiB | 0.0% (0/22 words) |
| small | LibriVox Gettysburg Address, 161.2 s | 465 MiB | 28.06 s | 5.74× | 11.10 s | 848 MiB | — |
| medium | JFK sample, 11.0 s | 1,463 MiB | 8.93 s | 1.23× | 4.36 s | 2,068 MiB | 0.0% (0/22 words) |
| medium | LibriVox Gettysburg Address, 161.2 s | 1,463 MiB | 60.64 s | 2.66× | 25.28 s | 2,171 MiB | — |
| large-v3-turbo q5_0 | JFK sample, 11.0 s | 547 MiB | 11.40 s | 0.96× | 9.89 s | 801 MiB | 0.0% (0/22 words) |
| large-v3-turbo q5_0 | LibriVox Gettysburg Address, 161.2 s | 547 MiB | 90.87 s | 1.77× | 77.45 s | 908 MiB | — |
| large-v3-turbo | JFK sample, 11.0 s | 1,549 MiB | 10.11 s | 1.09× | 7.77 s | 1,802 MiB | 0.0% (0/22 words) |
| large-v3-turbo | LibriVox Gettysburg Address, 161.2 s | 1,549 MiB | 74.38 s | 2.17× | 61.34 s | 1,909 MiB | — |
Method, commit, and the exact command: methodology.
On this old CPU, turbo was slower than medium. Turbo keeps large-v3’s full 32-layer encoder: the encoder was 61.34 s of turbo’s 74.38 s on the 161.2 s clip, and medium’s encoder was 25.28 s. OpenAI’s “~8×” for turbo is from an A100. The q5_0 file used less memory (908 MiB peak versus 1,909 MiB) and was slower here (90.87 s versus 74.38 s).
Every timed model got the 11.0 s JFK sample with 0 errors in 22 words. That clip is too easy to rank them. The .en models, large-v2, full large-v3, the Tesla P40, and any phone are Not yet measured.
Which one to pick
- English only, on a phone: base.en or small.en. OpenAI says the .en models tend to do better on English, mostly at tiny and base. Those two files are Not yet measured here.
- A first model with a page: base or small.
- Best accuracy for most people, on a GPU or Apple silicon: large-v3-turbo. On Frankenstein’s CPU, medium was faster. Turbo is large-v3 with the decoder cut from 32 layers to 4. It is the default in OpenAI’s own command-line tool.
- Translating another language into English: medium or large-v3. Turbo was not trained to translate. OpenAI’s README says it returns the original language even if you ask it to translate. Full large-v3 is Not yet measured on this machine.
- Just checking that a setup works: tiny. It was timed, and it does not have its own page.
Before you start
- OpenAI’s card says Whisper is strong in about 10 languages and uneven in the rest. Test it on your own recordings.
- It can repeat itself or make up text, especially in quiet stretches. The card says so.
- It reads audio in 30-second windows. Longer files work because the software slides along.
- In OpenAI’s Python package,
largenow means large-v3 andturbomeans large-v3-turbo. The original 2022 large is large-v1 and is not on this page. distil-whisper is a different family. - Whisper does not tell speakers apart. Apps add that with a separate model.
On a phone: Run Whisper on your iPhone. On a computer: Run Whisper on your computer. Ollama does not have an official Whisper model.
Wispr Flow is a popular dictation app, and it is not local. Its own data page says “Transcription always occurs on the cloud.” It uses Wispr’s model, not Whisper, and it does not get a model page. The note is on the iPhone guide.